easy-receptive-fields-pytorch

Minimal API for receptive field calculation in PyTorch

Requirements

• PyTorch >= 0.4
• Python 3
• MatPlotLib (optional)

Example

>>> import torch.nn as nn
>>> from receptivefield import receptivefield

>>> net = nn.Sequential(
  nn.Conv2d(in_channels=3, out_channels=32, stride=2, kernel_size=3),
  nn.ReLU(),
  nn.Conv2d(in_channels=32, out_channels=10, kernel_size=3),
)

>>> print(receptivefield(net, (1, 3, 32, 48)))

ReceptiveField(
  offset=Vector(x=0, y=0),
  stride=Vector(x=2, y=2),
  rfsize=Size(w=7, h=7),
  outputsize=Size(w=21, h=13),
  inputsize=Size(w=48, h=32)
)

Screenshot

ReceptiveField.show can be used to visualize the receptive field on a checkerboard image. Crosses denote the centers of receptive fields, and colored rectangles show their extent. Not all rectangles are plotted to avoid crowding the image.

Author

João F. Henriques